image credit: Martucci, Antonio (NRL) | accessed via http://www.fao.org
This week’s lab will explore following preprocessing steps necessary for working with multi-band satellite imagery:
During this week’s lab will utilize the following data:
Accessing Landsat Imagery:
Remote sensing analysis in desktop GIS requires input data from a satellite sensor. This imagery typically comes in a banded format whereby different wavelength values from the electromatic spectrum are distributed across different bands. These different bands are then combined to produce different visual and analysis products in desktop GIS.
A longstanding access interface for remote sensing imagery is EarthExplorer. Here a user creates a login profile and then has access to a range of products. Sometimes these products can be immediately downloaded, and sometimes they are prepared and then distributed via email notification.
Regarding the Landsat program which is the focus of this lab, there are several products that are available as noted in the table below:
Landsat Options
Landsat Interface at EarthExplorer
Typically level-1 products are more ‘raw’ than level-2 products. The level-1 products typically require more pre-processing in desktop GIS to produce the best results. The level-2 products are considered enhanced, but they usually require an ‘order’ and email notification as opposed to a direct download. The new (post-2018) ‘Analysis Ready Data’ (ARD) products are a highly tailored product designed to remove the burden of pre-processing. They also feature additional products outside the typical banded structure of level-1 and level-2. In this lab, we will utilize the ARD data to gain insight into recent burn areas in California, as well as explore ‘Land Surface Temperature’ (LST) in an urban environment.
The following steps outline navigation within EarthExplorer. For each component of the lab, data has been prepared from the EarthExplorer interface in order to save time. There are 5 essential steps to gaining access to imagery in EarthExplorer:
5 EarthExplorer Interface Components
EarthExplorer Results Options
To start, a Landsat 8 scence for Dubai, UAE has been ordered and placed in the following location for direct download - C11 Lab Data - Landsat Scene for Dubai, UAE.
LC08_L2SP_160043_20211012_20211019_02_T1
2021/10/12
WRS Path 160
WRS Row 043
Landsat Scene at EarthExplorer
Step 2 - Uncompress the Landsat product:
Download Keka to a Mac machine
.tar.gz..tar.gz file. The result = .tar file..tar.fileLandsat 8 Bands - https://www.harrisgeospatial.com
Dubia, UAE Landsat 8 Scene in Directory outside QGIS
STEP 3, open QGIS and point the Data Source Manager to the working directory as shown in the above image:
STEP 4, import just bands 1 - 7 Preview the individual bands in Layers Panel:
Individual Bands loaded to QGIS
Build Virtual Raster
Insert all 7 bands into the tool, and Toggle ON the Separate Bands option
The .vrt multispectal, multiband raster is now ready for analysis and/or band combinations. Typically the first step is to develop an appropriate band combination to accentuate a certain visual characteristic and/or land cover feature. Importantly, each satellite product has its own band combinations. Listed here are typical combinations of the Landsat 8 product. To finish this Part II lab, 3 band combinations will be applied and then viewed:
Natural Color (4, 3, 2)
Color Infrared (5, 4, 3)
Short-Wave Infrared (7, 6 4)
Agriculture (6, 5, 2)
Geology (7, 6, 2)
More details on band combinations located HERE
To enact a band combination, navigate to Layer Properties > Symbology > Multiband Color and set the first combination to 4-3-2.
4-3-2 Band Combination
This band combination is known as ‘True Color’ as the band 4 is the red color band, 3 is the green color band and 2 is the blue color band. This is the color range we typically ‘see’, i.e. looking out an airplane window. Any other band combination is a ‘False Color’ meaning we are forcing wavelength values into the red/green/blue color space.
To start the first band combination 4-3-2 set symbology as follows. Notice that the image looks washed out and hazy. This is because the current image is stretched across all values in the raster. This can be corrected two ways. First, set the cumulative cut upwards towards 10% instead of 2%; and secondly, set the statistical extent to the map canvas. Going forward, the image will render much more accurately:
Cumulative Cut
Raster Extent
4/3/2 True Color
7/6/4 False Color
5/4/3 CIR Infrared
In Class 11 assignment, following the band creation process for Landsat imagery, the imagery will be trained and classed. We will review this process during both this lab and the lecture demo.
The imagery will be classified into ‘classes’, ‘bins’ or ‘buckets’ of similar pixels that are meaningful for a particular analysis. In the assignment, we will be interested in just 2 classes - water and land.
We will utilize the dzetsaka tool plugin for the lecture demo, lab and assignment -this will be the classification tool for this week.
There are several cluster algorithms in the tool that are available. However, depending on your machine and setup, you may need to download the scikit-learn python library in QGIS. Directions for this step are located at the assignment 11 towards the end. You will only need to do this installation if you are:
Gaussian Mixture algorithm.k-Means Neighbors method.demo_dzetsaka.qgs:dzetsaka tool sample data
dzetsaka tool plugin
Training Shapfile + Raster Map
Here a old map has been rasterized and georeferenced into QGIS. From this map, the 5 features in the training correspond with locations in the old map where those land types exist. The classification will go through the image, and determine all areas within the raster that fit the pattern of each of the 5 training features.
Populate the tool as follows:
Class will be used as the training input for this demonstration
Gaussian Mixture method first:Gaussian Mixture
k-Means Neighbors method:k-Means Neighbors
Methods Comparison
k-Means Neighbors takes significantly longer to run than Gaussian Mixture. Importantly, k-Means Neighbors only picks up 4 classes, whereas Gaussian Mixture picks up all five. If the training samples are reviewed, its clear that the forest class crosses a red road; in the k-Means Neighbors method it lumps this read into the forest class, and disregards the 1 polygon in the training for Buildings.